HHAvatar: Gaussian Head Avatar with Dynamic Hairs

Area of Science:

• Computer Vision
• Computer Graphics
• 3D Reconstruction

Background:

• Creating high-fidelity 3D head avatars is challenging, especially with limited camera views (sparse views).
• Existing methods struggle with dynamic details like hair and accurate facial expressions under sparse conditions.

Purpose of the Study:

• To propose HHAvatar, a novel method for high-fidelity 3D head avatar creation using controllable 3D Gaussians.
• To enable dynamic hair modeling and accurate expression capture from sparse view setups.

Main Methods:

• Representing head appearance with 3D Gaussians and optimizing them with a learned MLP-based deformation field for expressions.
• Employing a geometry-guided initialization strategy using implicit Signed Distance Fields (SDF) and Deep Marching Tetrahedra for stable training.
• Introducing a hybrid head model and a specialized training method with timing and occlusion perception for dynamic hair modeling.

Main Results:

• HHAvatar achieves ultra high-fidelity rendering quality at 2K resolution, outperforming state-of-the-art sparse-view methods.
• The method successfully models fine-grained dynamic details and ensures expression accuracy.
• Dynamic hair motion is reasonably driven by head movements, even under exaggerated expressions.

Conclusions:

• HHAvatar offers a robust solution for high-fidelity 3D head avatar creation from sparse views.
• The integration of 3D Gaussians, deformation fields, and dynamic hair modeling advances the field.